Case Studies

OpenAI Execs Knew About Pirated Books Used to Train GPT-4

Unsealed court documents show OpenAI leadership proceeded with GPT-4 training despite internal warnings about the pirated Books3 dataset. The real debate among engineers isn't whether the copying was wrong—it's who ends up paying for it.

Font size

What the Complaint Actually Says: More Than "We Found It Online"

In late August, the key complaint in Authors Guild v. OpenAI was unsealed. The Authors Guild published a piece that same day titled AG v OpenAI: Top Execs Knew Mass Book Piracy Was Illegal—a headline that tells you where the argument lands.

The complaint isn't about the fact that OpenAI used pirated books. That's been an open industry secret for years. What it alleges is a timing problem:

  • OpenAI insiders—including Sam Altman and Brad Lightcap—received repeated warnings that the books came from pirated sources
  • Internal emails and Slack records show employees explicitly noting that Books3's data lineage traced back to Library Genesis, Anna's Archive, and similar shadow libraries
  • Despite these warnings, OpenAI pushed the data through the GPT-4 training pipeline anyway

OpenAI's natural defense would have been "we didn't know it was pirated"—an innocent infringement claim. The complaint instead argues willful infringement: you knew, and you kept going.

These two framings differ in damages by an order of magnitude. Under 17 U.S.C. § 504(c)(2), statutory damages for willful infringement cap at $150,000 per work. With Books3 weighing in at roughly 196,000 books and 170 GB of text, the theoretical ceiling is astronomical. Courts rarely award the maximum—but the fact that "the CEO knew" alone pushes settlement figures into the eight-figure range.

Why This Isn't Just an OpenAI Problem

If the story were only about OpenAI's fine, Hacker News wouldn't have given it 533 points.

What got the engineering crowd fired up: Books3 wasn't exclusive to OpenAI.

  • Meta's LLaMA series openly acknowledged Books3 as one of its training sources
  • Mistral's earlier models also ingested the dataset
  • Anthropic took a "licensed procurement" route for Claude training, but reports from August 2023 indicated they used Books3 derivatives
  • Hugging Face still hosts various "cleaned" and "deduplicated" derivatives of Books3

One passage in the complaint hits particularly hard: the claim that OpenAI's conduct isn't an outlier but the industry standard. If the fair-use defense collapses, every data licensing contract in the industry needs to be rewritten.

That's why employees from Meta and Microsoft were also following the thread. They want to know: is my company next?

What Engineers Are Arguing About on HN

The original HN thread pulled 533 points and 468 comments—well above the usual engagement for a copyright story. The split isn't about whether OpenAI was wrong to take the books. Almost nobody defends that. The real question is:

If someone's paying, who is it?

A few top-voted comments:

"If Microsoft is a co-defendant, does that make Azure an accomplice? They provided the compute." (+247)

"I uploaded a deduplicated version of Books3 to HF. Should I delete it now?" (+189)

"The $150k/work statutory cap is the ceiling—real cases usually land at 1/10 to 1/100 of that. But at OpenAI's scale, their legal fees alone could cover buying every book legitimately." (+134)

"The real question: post-2026, new models need full data provenance logs. Anthropic is already doing this. OpenAI is still playing dumb." (+98)

Engineers aren't worried about whether AI companies will be destroyed by fines. They're worried about whether they've personally crossed a line, whether the models they're using have clean data sources, and how to write their next contract.

Practical Impact for Developers Outside the US

The fallout lands in three layers for developers in mainland China, Hong Kong, and Taiwan:

Layer 1: Model selection

If you're currently using Qwen, DeepSeek, Yi, or other domestic models, each vendor's model card includes a "training data is legal and compliant" line. The specifics of how that compliance was achieved—whether Chinese-language pirated book sources (the Chinese corners of z-library, various cloud-drive libraries) were filtered out—haven't been disclosed.

If this ruling establishes "provenance must be auditable" as an industry standard, those one-line compliance declarations won't cut it. Auditable data-source manifests will be required.

Layer 2: API costs

The complaint cites the $150,000-per-work statutory ceiling. If a settlement lands in the $1B–$5B range (roughly the zone recent Anthropic copyright settlements have been speculated to fall in), that cost will eventually pass through to API pricing.

OpenAI has already raised o1 API prices once. If a settlement lands on top of that, GPT-5 / Claude 5 API prices in 2026 could rise another 20–30%.

Layer 3: Domestic data procurement

The mainstream Chinese "DataTang-style" model—buying a batch of cleaned datasets with broad usage rights—will be judged against this ruling. If U.S. courts find that "knowing use" constitutes willful infringement, the liability disclaimers in those Chinese data-supplier contracts likely won't hold up.

Teams already running RAG or fine-tuning on Chinese text should check whether their pretraining data has any traceable provenance chain. Nobody was auditing this before. They will be soon.

Editor's Take

What makes this story worth watching isn't the copyright question itself—it's the possibility of an "Enron moment" for the AI industry. Internal emails getting unsealed makes willful infringement stick in a way that the "we didn't know" defense never could.

Two predictions. First, the case will almost certainly end in a monetary settlement rather than a ruling that overturns generative AI training. But it will require every new model trained after 2026 to leave a complete provenance trail on its pretraining data. Anthropic is already doing this.

Second, if Chinese vendors keep walking the "compliant but not detailed" line, this ruling becomes a frequently-cited counter-example.

What sets this apart from earlier copyright cases: NYT v. OpenAI and Getty v. Stability targeted "similar outputs." This one targets "known inputs." The former can be addressed with technical filtering (memorization mitigation). The latter requires restructuring procurement workflows. Once a court accepts the willful infringement framing, data procurement becomes a permanent item on the compliance checklist—a constraint the industry has never operated under before.

What I'm uncertain about: whether U.S. courts' recognition of "industry custom" will extend to non-U.S. vendors. If it doesn't, the Chinese "DataTang" model faces no short-term pressure. If it does, Qwen and DeepSeek's overseas deployments are immediately exposed—worth tracking as the ruling develops.

# Sources

https://authorsguild.org/news/ag-v-openai-top-execs-knew-mass-book-piracy-was-illegal/

Source: https://authorsguild.org/news/ag-v-openai-top-execs-knew-mass-book-piracy-was-illegal/